Geography is modeled directly
State–PUMA has the largest global mean absolute SHAP value (0.344). It is shown separately so it does not obscure the substantive predictors.
Interactive research · ACS PUMS 2020–2024
Geographic variation in the drivers of predicted housing value
Explainable machine learning across California, Florida, New York, Tennessee, and Texas—tracking how the same fitted model relies on housing, financial, household, and geographic information.
5states
960.2Kdevelopment records
2020–23model development
2024temporal evaluation
268.9Krecords explained
01 · Research landscape
One study, five housing markets
The 2024 explanation set spans 268,930 owner-occupied one-family housing records. Survey-weighted aggregation keeps the descriptive and geographic summaries aligned with the study design.
02 · Model explanations
What the selected model relies on
Geography is prominent, but mortgage payment, household income, property age, and room configuration supply the leading substantive signals.
Survey-weighted mean absolute TreeSHAP for 2024 records. Larger values mean the selected model relied more on the feature; units are not dollars.
Mortgage and income lead beyond geography
First-mortgage payment and household income are the first two non-geographic features in every state-level ranking.
03 · Geographic heterogeneity
The ranking is shared. The reliance is not identical.
State summaries reveal variation in how strongly the fitted model uses the same non-geographic features.
FeatureCaliforniaTennessee
First mortgage payment0.1750.140
Household income0.0850.085
Year built0.0330.077
For example, year built carries more than twice the mean absolute SHAP importance in Tennessee (0.077) than in California (0.033). This is variation in model explanation—not evidence of different causal effects.
04 · Validation
Selected before the clock moved forward
The 2024 data are not another random holdout. All model comparison and feature reduction occurred within 2020–2023 development folds; 2024 provides a temporal test of generalization.
- Mean development CV MAE
- $247,093
- 2024 temporal MAE
- $275,639
- 2024 median absolute error
- $113,951
- 2024 R²
- 0.318
2020–2023 mean CV MAE2024 temporal MAE